DX 12 collective memory

[atmos89]

I'm just curious what DX 12 means for octane and if octane V3 will have implemented support for it and its stacking memory feature. Will it be possible to cumulate Vram, will octane see all the gpu's vram as one big pool ?

[grimm]

No, Octane doesn't have anything to do with DX12 as it uses CUDA and soon OpenCL to do the image rendering. DX 12 is Windows only as well so it wouldn't work at all for Macs or Linux boxes.

[atmos89]

so even the future NV-link won't work for octane ?

[grimm]

I think that, if it helps with rendering, the developers could use it but I don't know enough about NVlink to say for sure. It looks like it will be unlikely that consumer GPUs will get NVlink any time soon and only Tesla cards have it right now. Tesla cards are very expensive and not overly useful for GPU rendering like Quadro cards (except having lots and lots of memory). Tesla's are used in big supercomputers and the like for their double precision floating point calculations. Most GPU renderers use single precision floating point calculations because the current consumer cards do it very quickly. We can thank games for that.

[smicha]

I think that, if it helps with rendering, the developers could use it but I don't know enough about NVlink to say for sure. It looks like it will be unlikely that consumer GPUs will get NVlink any time soon and only Tesla cards have it right now. Tesla cards are very expensive and not overly useful for GPU rendering like Quadro cards (except having lots and lots of memory). Tesla's are used in big supercomputers and the like for their double precision floating point calculations. Most GPU renderers use single precision floating point calculations because the current consumer cards do it very quickly. We can thank games for that.

Grimm,

I cannot find it anywhere - will nvlink replace PCIe slot physically? I mean will pascal require new motherboards?
From this video I cannot judge it
http://www.nvidia.com/object/nvlink.html

[grimm]

Hi Smicha,

Sorry I didn't respond sooner, work was a bear today. I took a quick look at the white paper and it looks like NVlink is mostly a faster and wider SLI connection. It's kind of confusing actually as the white paper starts off saying that it's a faster link between the GPU and the CPU but ends up just talking about linking GPUs together. According to the white paper they are planning to implement it with Pascal. Each card will have 4 NVlinks that each can talk up to 20Gbits per second. You can stack the links together to get up to 80 Gbits per second depending on how many cards you want to interconnect (max 4). For those applications where you need to share data between GPUs (games?) it might help a lot. For rendering I'm not so sure?

[smicha]

Thanks Grimm. That is how I understood it. So PCI slots are still valid, aren't they?

[grimm]

Looks that way, at least short term. It looks like PCI is going to be used for SSD and disk connections more and more (m.2). I don't know what that will mean for GPUs and I haven't really seen a good replacement for PCI yet.

[atmos89]

can vulcan api help memory pooling ?